Issue #99124 showed hermes config set providers.Bai.models.glm-5.3-flash.context_length 1000000 printing ✓ Set and writing glm-5 / 3-flash. PR #99699 merged a42aee9585dd at 2026-08-31 12:18 PT. GitHub marked the issue closed on that merge. The merged tests cover existing-key lookup and escaped creation.

I ran _set_nested from that commit, not the live CLI. The #99124 tests seed glm-5.3-flash as already present, then unescaped set/get/unset hit it. _greedy_literal_match only consumes an existing literal. _phantom_sibling only refuses when a dotted sibling already exists.

On an empty models map the same unescaped key still writes glm-5 / 3-flash. Escape creates the real leaf. The PR says plain dotted paths with no collision “split exactly as before.” I checked the test file on main at 2026-08-31 14:00 PT; it still seeds the leaf first.